Circuit OS: AI Hardware Compiler on AMD ROCm

Circuit OS is an open-source (AGPL-3.0) AI hardware compiler: describe a circuit in plain English and get a validated schematic, Arduino firmware, a SPICE simulation, a bill of materials, and a plain-English explanation of every design choice. Built for AMD: the AI layer runs on open models served by vLLM on AMD ROCm GPUs (AMD Developer Cloud), as well as on Claude. One setting (AI_PROVIDER=openai_compat) switches providers, and scripts/amd_benchmark.py measures valid-design rate, retries and latency on the GPU. The core rule is that the AI never writes SPICE, KiCad files or firmware directly. LLMs invent pin numbers and function names, so the model only outputs JSON against a strict intermediate representation (IR) schema through forced tool calls (guided decoding on vLLM). Pydantic and a hardware rule engine check the IR for floating nodes, voltage ratings, missing I2C pull-ups, RS-485 termination and PWM-capable pins. If validation fails, the model gets the exact field errors and retries up to three times, which is what makes smaller open models usable. Deterministic compilers then turn the IR into an ngspice netlist, a KiCad schematic, Jinja2 Arduino code and a BOM. Same input, same output. Simulations run in the background via Celery and Redis and are graded pass/fail with a 15% tolerance, so a wrong part (1nF instead of 100nF in an RC filter) is caught before anything is built. Follow-ups like "make the LED brighter" become small patches that change only what was asked. Explanations focus on consequences: what breaks if you change R3. v0.1.0: five templates (DHT22 alert, MAX485 Modbus RTU, LED, RC low-pass filter, voltage divider), JWT auth, rate limiting, PostgreSQL, 430+ passing tests, firmware that compiles for a real Arduino Uno, and an experimental A* PCB router. Next: free-form circuits and fine-tuning an open model on AMD GPUs.